Jumping champions and prime gaps using information-theoretic tools Nicholas Pun,1 Robert T.W. Martin,2 and Achim Kempf3, 4 1)Department of Applied Mathematics, University of Waterloo 2)Department of Mathematics and Applied Mathematics, University of Cape Town 3)Departments of Applied Mathematics and Physics, University of Waterloo 4)Institute for Quantum Computing, University of Waterloo (Dated: 3 August 2018) We study the spacing of the primes using methods from information theory. In information theory, the equivalence of continuous and discrete representations of information is established by Shannon sampling theory. Here, we use Shannon sam- pling methods to construct continuous functions whose varying bandwidth follows the distribution of the prime numbers. The Fourier transforms of these signals spike at frequently occurring spacings between the primes. We find prominent spikes, in particular, at the primorials. Previously, the primorials have been conjectured to be the most frequent gaps between subsequent primes, the so-called “jumping cham- pions”. Here, we find a foreshadowing of the primorial’s role as jumping champions in the sense that Fourier spikes for the primorials arise much earlier on the num- ber axis than where the primorials in question are expected to reign as jumping champions.

I. INTRODUCTION

The gaps between the primes possess intriguing structural properties and have led to a number of important as yet unproven conjectures, such as the Hardy-Littlewood k-tuple conjecture1 of 1923. More recently, in 1999, Odlyzko, Rubinstein and Wolf2 published a conjecture concerning so-called jumping champions. For any t > 0, the jumping champion is defined as the integer, g, that is the most frequently occurring gap between any two successive primes less than or equal to t. The jumping champions conjecture states: Conjecture 1 The jumping champions are 4 and the primorials, i.e., 2, 6, 30, 210, .... Here, the n’th primorial is the product of the first n primes. In 2012, Goldston and Ledoan proved that a version of the Hardy-Littlewood k-tuple conjeture for prime pairs and triples implies that all sufficiently large jumping champions are primorials, and any sufficiently large primorials are jumping champions over a long range of t > 0, see3. In particular, they provide estimates on ranges of t for which a given primorial is the jumping champion for [0, t]. For example, the primorial 210 is expected to reign as jumping champion in (roughly) the interval [10487, 102607]. See4–7 for related results and investigations. arXiv:1808.00572v1 [math.NT] 1 Aug 2018 The magnitude of these numbers would appear to preclude numerical studies. However, as we will show, intriguing evidence for the importance of the primorials in the distribution of prime gaps can be obtained numerically through the use of the information-theoretic tools of Shannon sampling theory. In information theory, Shannon sampling constructively estab- lishes the equivalence between discrete and continuous representations of information8–12. Our aim here is to use Shannon sampling methods to map the discrete structure given by the primes into continuous functions which can then be Fourier analyzed. Concretely, our study has three parts. In the first part, we consider a histogram of the spacings between any pair of primes within some finite interval. In the second part, we use the primes to construct a continuous function by using Shannon sampling theory which is then Fourier analyzed. In the third part, we apply a generalized Shannon sampling method. With each method, we find that the primorial’s role as jumping champions is foreshadowed in the sense that Fourier spikes for the primorials arise much earlier on the number axis than where the primorials in question are expected to reign as jumping champion. In addition, we also find prominent Fourier spikes at frequencies that would correspond to 2 certain non-integer spacings. These spacing are simple ratios that are as yet unexplained.

II. HISTOGRAM ANALYSIS OF PRIME GAPS

We begin our analysis of the prime gaps by plotting a histogram of the differences between consecutive primes up to a maximum prime (Figure 1).

FIG. 1. Distribution of the differences between consecutive primes, up to the 50000th prime

Previously, Ares and Castro13 noticed that periodic oscillations occur within the his- togram, leading to spikes at differences that are a multiple of 6. Motivated by this observa- tion, let us now examine if similar structural properties exist among the differences between primes that need not be consecutive, see (Figure 2).

FIG. 2. Distribution of the differences between every combination of primes, beginning from the second prime (p2 = 3) to the 50000th prime. We find that some of the prominent spikes are at distance values that “foreshadow” the occurrence of jumping champions.

Figure 2 exhibits a significant number of spikes and a closer inspection shows that among them are all the primorials up until the primorial 2310. This very early occurrence of 3 such large primorials suggest closer examination with more powerful tools. To this end, we will now employ Shannon sampling methods from information-theory. We will use both regular and adaptive Shannon methods to map the discrete distribution of the primes into continuous functions whose periodicities, which are related to spacings of primes, can then be analyzed using the Fourier transform.

III. FREQUENCY ANALYSIS USING CLASSICAL SHANNON SAMPLING THEORY

In information theory, the Shannon sampling theorem plays a central role as it establishes an equivalence of discrete and continuous representations of information. Concretely, it allows one to perfectly reconstruct a bandlimited (and therefore continuous) function on the real line from knowledge of its amplitudes only at a discrete set of points on the real line. We recall that a function over the reals is called bandlimited if the support of its Fourier transform is bounded. We will use the Shannon sampling theorem to construct a continuous and bandlimited function by specifying its samples on the integers to be either 1 or 0 depending on whether the integer is prime or not. We then Fourier analyze the constructed function.

A. Background

The Shannon sampling theorem8–12 states that if a function, or ‘signal’, φ(t), possesses no frequencies above some finite value Ωmax, then it suffices to record the samples {φ(tn)} −1 on an equidistantly-spaced lattice {tn} with spacing tn+1 − tn = (2Ωmax) . If the samples are taken at this rate, the so-called Nyquist Rate, the function φ(t) can be reconstructed for all t through

∞ X φ(t) = G(t, tn)φ(tn) (III.1) n=−∞ where G(t, tn), the reconstruction kernel, is defined as:

G(t, tn) = sinc (2(t − tn)Ωmax) (III.2)

We remark that Ωmax-bandlimited functions can also be reconstructed from non-equidistantly spaced samples, if their average density (in the sense of Beurling) matches or exceeds the Nyquist density. The reconstruction from a non-equidistantly-spaced sampling lattice is necessarily less stable, however, in the sense that small measurement errors in the ampli- tudes translate into increased errors in the reconstructed function.

B. Signal Construction and Analysis

Our aim is to construct a continuous function, or ‘signal’, Φ(t), based on the primes. To this end, we define our sampling lattice {tn} to be the set of integers and we define the function’s amplitudes on the integers to be: ( 0, if tn non-prime φ(tn) = 1, if tn prime

The resulting signal obtained by using the first 50000 primes is shown in Figure 3 and its Fourier spectrum is shown in Figure 4. 4

FIG. 3. Zoom-In of prime number signal generated using the Shannon sampling theorem

FIG. 4. Fourier transform of Prime Signal Generated Using Classical Reconstruction Methods.Only the positive frequencies are shown, and the leftmost frequencies correspond to longer wavelengths. The most prominent spikes are dotted.

No. Frequency (Hz) Amplitude Wavelength 1 673148 4941.04 20.000 2 713945 2241.09 12.000 3 734343 2252.28 6.6667 4 795539 4766.72 6.0000 5 856734 2490.08 5.0000 6 917929 5591.44 4.0000

TABLE I. The frequency, amplitude and wavelength of the most prominent spikes, from left to right in Figure 4. 5

In Fig.4 and Table 1, the occurrence of spikes at certain integer wavelengths indicates the prevalence of corresponding prime gaps. In particular, 6 appears among those with highest amplitude, indicating that it is one of the most commonly occurring prime gaps, consistent with the observations made in Section II. We notice also that there are spikes at non-integer wavelengths which appear to be simple ratios. These non-integer ‘effective gaps’ indicate the existence of structures that cannot be seen in a histogram of integer prime distances.

IV. FREQUENCY ANALYSIS USING GENERALIZED, ADAPTIVE SHANNON SAMPLING THEORY

Our aim now is to try to further amplify the phenomenon of the foreshadowing of jumping champions and the occurrence of non-integer prominent prime spacings by applying a gen- eralized Shannon method that allows one to adapt the choice of sample points to arbitrary irregular spacings. This allows us to choose our sample points to be the sequence of the primes. We will again set the amplitude at the sample points to be 1 to obtain a signal that can then be Fourier analyzed. Notice that if we also require the amplitude to be 0 at the non-prime integers, then we recover our previous signal exactly. So, our new method is distinguished from the method of the previous section by in this sense only focusing on the prime numbers.

A. Background

The generalized sampling theory14–21 generalizes the regular Shannon sampling theorem of functions, or signals, that possess a constant bandwidth and constant Nyquist rate to classes of functions that possess a time-varying bandwidth, or, correspondingly, a time- varying Nyquist rate. This allows one to consider classes of signals of time-varying band- width that can be most stably reconstructed from their amplitudes on a sampling lattice, {tn}, whose spacing correspondingly varies in time. In addition to specifying the sampling lattice {tn}, which we will choose to be the primes, the generalized sampling theory also 0 requires the specification of a set of values {tn} which in effect describe the extent to which the bandwidth may change from sample to sample. In the absence of additional information 0 that we could use here, we will set these values to be the standard values of tn = tn+1 − tn. The reconstruction formula Eq. (III.1) can now be applied with the generalized recon- struction kernel14–21: p !2 t0 X t0 G(t, t ) = (−1)z(t,tn) n m (IV.1) n t − t (t − t )2 n m m

The function z(t, tn), in the exponent is the number of sampling points between t and tn, z so that (−1) (t, tn) makes G(t, tn) differentiable. More generally, given s, t ∈ R, X t0  G(s, t) := f(t) · n · f(s), (t − tn)(s − tn) where t0 z(t,tn) −1/2 X n f(t) := (−1) g(t) ; g(t) := 2 , (t − tn) is a smooth (infinitely differentiable) positive kernel function on R × R in the sense of reproducing kernel Hilbert space (RKHS) theory, and our space of functions obeying a ‘generalized’ bandlimit is the unique RKHS H(G) corresponding to G, see20 (Section 2). The theory of these spaces is closely connected to the theory of Hardy spaces of analytic 6 functions in the complex upper-half plane20,22–31: Given any generalized RKHS, H(G), of time-varying or locally bandlimited functions, one can find a fixed function, M(t), so that multiplication by M(t) is a unitary transformation of H(G) onto a co-invariant (for the shift) subspace of the Hardy space of the upper half-plane which is the orthogonal complement of the range of a meromorphic inner function20. This inner function can be expressed explicitly 0 20 in terms of the sequences of sample points {tn}, and their ‘speeds’ {tn} . The classical Shannon sampling kernel can be recovered as a special case of the generalized kernel G(s, t) with the choice of sampling sequences: nπ π t := , and t0 := tanh(A) ; n ∈ . n A n A Z Indeed, one can apply trigonometric series identities to show

1 X  1 1  − = cot(πt) − cot(πs), π t − k s − k k∈Z

g(t)−2 = Aπ tanh(A) csc2(At), and finally that

sin ((t − s)A) G(s, t) = , (t − s)A see15 (Section 8.2) or20 (Example 2.28).

B. Signal Construction and Analysis

We construct our signal φ(t) as follows: our sampling lattice {tn} is a finite set of consec- 0 utive prime numbers, and φ(ti) = 1 for all i. Further, we set {tn} = { (ti+1 − ti−1)/2 | i = 0 0 2, . . . , n − 1 } and t1 = (t2 − t1)/2 and tn = (tn − tn−1)/2. As with our previous signal, we choose our sampling lattice to be the first 50000 primes, and the results of our reconstructed signal can be seen below.

FIG. 5. Zoom-in of the prime signal containing 50000 primes. The dotted points highlight where the prime numbers lie in the signal. 7

FIG. 6. Modulus of the Fourier transform of the signal in Figure 5.

No. Frequency (Hz) Amplitude Wavelength 1 2915 2146 210.02 2 20399 2981 30.001 3 27817 3438 22.000 4 43712 4306 14.000 5 61196 3377 10.000 6 101993 2837 6.0000 7 142790 1466 4.2857 8 203985 816.9 3.0000 9 244782 336.2 2.5000 10 265181 186.69 2.3077

TABLE II. Calculated wavelengths from the Fourier transform in Figure 6. Notice that the non- integer spacing 4.2857 can be written as 30/7.

The result of using generalized sampling theory, compared to the previous two methods, is that we have uncovered the same and even more structure in the prime gaps. Previous results, such as the common gaps at multiples of 6, or the non-integer values are reproduced using this method. In addition, we find a foreshadowing of the next two conjectured jumping champions, 30 and 210. Given Conjecture 1), this suggests that the generalized sampling method is indeed able to accurately highlight frequent prime gaps. Let us now recall that in 2012, Goldon and Ledoan3 provided several intervals in which a primorial is likely to become a jumping champion. The intervals are shown in Table III, # where pk denotes the product of the first k primes (the k-th primorial).

# No. k pk Interval 1 2 6 [4.67 ∗ 104, 2.32 ∗ 108] 2 3 30 [2.06 ∗ 1044, 5.24 ∗ 10150] 3 4 210 [4.64 ∗ 10487, 4.01 ∗ 102607] 4 5 2310 [8.78 ∗ 107769, 1.72 ∗ 1060178] 5 6 30030 [9.70 ∗ 10134460, 1.72 ∗ 101386286]

# TABLE III. Intervals in which pk is likely to be a Jumping Champion

These numbers indicate that in order to confirm 30 as a jumping champion it is necessary 8 to study the prime numbers up to about 1044. Yet, by creating a signal from only 50000 primes, a spike corresponding to the jumping champion 30 can already be observed in the Fourier transform. Likewise, seeing 210 foreshadowed as a spike is surprising. A possible explanation for the early occurrence of the primorials is that the primorials may not only be the jumping champions of subsequent primes but may also be closely related to the most frequently occurring gaps between any pairs of primes.

V. STABILITY ANALYSIS

A. Comparison of the distribution of primes to a Poisson-distributed sequence of integers

For comparison, we now apply our method also to Poisson-distributed sets of integers of the same density as the primes, see Figure 7. We observe in Figure 8 that there is an accumulation of low frequencies in the Poisson case which in the case of the prime numbers become re-distributed to become the prominent spikes.

FIG. 7. Top: Section of signal constructed from the first 50000 primes, Bottom: Section of signal constructed from 50000 Poisson-distributed numbers 9

FIG. 8. Top: Fourier transform of prime signal, Bottom: Fourier transform of poisson-distributed signal

B. Verification of the translation-invariance of the generalized sampling methods

Here, we apply our methods to three sampling lattices from various sections of the prime numbers: the 1st to 10000th prime, the 10001st to 20000th prime, and the 20001st to 30000th prime. We then calculate the Fourier transforms of the signals constructed from these sampling lattices. We observe that the prominent spikes are stable in the sense that they are prominent in all three Fourier transforms.

FIG. 9. Top: Fourier transform of signal constructed from the 1st to 10000th prime Center: Fourier transform of signal constructed from the 10001st to 20000th prime Bottom: Fourier transform of signal constructed from the 20001st to 30000th prime 10

VI. CONCLUSIONS AND OUTLOOK

We analyzed the discrete distribution of the primes by Fourier analyzing continuous func- tions obtained from the primes. In order to map the sequence of primes into continuous functions, we used Shannon sampling methods from information theory. The local behav- ior of the information-theoretically obtained continuous function depends on the primes nonlocally, with the influence of primes a distance of d away naturally decaying as 1/d. The Fourier transform of these continuous ‘prime signals’ yielded intriguing peaks at the primorials, as well as at non-integer wavelengths. In particular, the application of adaptive Shannon methods yielded more and stronger peaks at the primorials. We conclude that, for as yet unknown reasons, the presence of the jumping champions of Conjecture 1, is foreshadowed far earlier on the number line, among very much smaller primes than expected. This suggests that the primorials may also play a prominent role for the distances among any two not necessarily subsequent primes, giving rise to long-range correlations. Also, intriguingly, our results show prominent wavelengths in the Fourier analysis of the prime signals that occur at values that are not integer and that therefore cannot directly correspond to prime gaps. These wavelengths, which may be called effective prime gaps, appear to be particularly simple ratios whose origin and structure should be very interesting to explore, as they may be related to the Chebychev bias in the distribution of primes, see, e.g.32, or more generally to the biases that were recently discovered in33.

Application of generalized Shannon sampling method to other sequences

It should also be very interesting to apply the new method that uses adaptive Shan- non sampling to other sequences. For example, we have applied the new method to the sequence of squares and twice the squares (SEQ1), and the sequence of integers that are the sums of two squares (SEQ2), as shown in the figures below.

FIG. 10. Top: Signal constructed of numbers from SEQ1 less than 209760. Bottom: Signal constructed from Poisson-distributed numbers of the same density as SEQ1 11

FIG. 11. Top: Fourier transform of signal constructed from SEQ1. Bottom: Fourier transform of Poisson-distributed numbers of the same density as SEQ1

FIG. 12. Top: Fourier transform of signal constructed from SEQ1 with a logarithmic scale for the x-axis Bottom: Fourier transform of Poisson-distributed numbers of the same density as SEQ1 with a logarithmic scale for the x-axis 12

FIG. 13. Top: Signal constructed from the first 5871 integers of SEQ2. Bottom: Signal constructed from Poisson-distributed numbers of the same density as SEQ2

FIG. 14. Top: Fourier transform of signal constructed from SEQ2. Bottom: Fourier transform of Poisson-distributed numbers of the same density as SEQ2

Acknowledgements: NP, AK and RTW are grateful for useful feedback from Kevin Hare, Tristan Freiberg and Stefan Steinerberger. AK acknowledges support from the Discovery program of the National Science and Engineering Research Council of Canada (NSERC). 13

REFERENCES

1G. H. Hardy and J. E. Littlewood, “Some problems of partitio numerorum; iii: On the expression of a number as a sum of primes.” Acta Math. 44, 1–70 (1923). 2A. Odlyzko, M. Rubinstein, and M. Wolf, “Jumping champions,” Experimental Mathematics 8, 107–118 (1999). 3D. Goldston and A. Ledoan, “The jumping champion conjecture,” Mathematika 61, 719–740 (2015), arXiv:1102.4879v2 [math.NT]. 4M. Wolf, “Applications of statistical mechanics in number theory,” Physica A: Statistical Mechanics and its Applications 274, 149–157 (1999). 5M. Wolf, “Some heuristics on the gaps between consecutive primes,” arXiv preprint arXiv:1102.0481 (2011). 6S. Ares and M. Castro, “Hidden structure in the randomness of the prime number sequence?” Physica A: Statistical Mechanics and its Applications 360, 285–296 (2006). 7G. Szpiro, “The gaps between the gaps: some patterns in the prime number sequence,” Physica A: Statistical Mechanics and its Applications 341, 607–617 (2004). 8C. E. Shannon, “Communication in the presence of noise,” Proc. IEEE 86, 447–457 (1998). 9A. Jerri, “The Shannon sampling theorem - its various extensions and applications: A tutorial review,” Proc. IEEE 65, 1565–1596 (1977). 10J. Benedetto and P. Ferreira, Modern sampling theory: mathematics and applications (Springer, 2012). 11A. I. Zayed, Advances in Shannon’s Sampling Theory (CRC Press, Inc., 1993). 12R. Marks, Introduction to Shannon sampling and interpolation theory (Springer, 2012). 13S. Ares and M. Castro, “Hidden structure in the randomness of the prime number sequence?” Physica A 360, 285–296 (2006), arXiv:0310148v2 [cond-mat]. 14A. Kempf, “Fields over unsharp coordinates,” Phys. Rev. Lett. 85, 2873 (2000). 15A. Kempf, “Fields with finite information density,” Phys. Rev. D 69, 124014 (2004). 16Y. Hao and A. Kempf, “Generalized Shannon sampling method reduces the Gibbs overshoot in the approximation of a step function,” J. Concr. Appl. Math. 8, 540–554 (2010). 17Y. Hao and A. Kempf, “On a non-fourier generalization of shannon sampling theory,” in Information Theory, 2007. CWIT ’07. 10th Canadian Workshop on (2007) pp. 193–196. 18Y. Hao and A. Kempf, “On the stability of a generalized shannon sampling theorem,” in Information Theory and Its Applications, 2008. ISITA 2008. International Symposium on (2008) pp. 1–6. 19Y. Hao and A. Kempf, “Filtering, sampling, and reconstruction with time-varying bandwidths,” IEEE Signal Proc. Lett. 17, 241–244 (2010). 20R. Martin and A. Kempf, “Function spaces obeying a time-varying bandlimit,” J. Math. Anal. Appl. 458, 1597–1638 (2018). 21Yufang Hao, Generalizing Sampling Theory for Time-Varying Nyquist Rates using Self-Adjoint Exten- sions of Symmetric Operators with Deficiency Indices (1,1) in Hilbert Spaces, Ph.D. thesis, University of Waterloo (2011), retrieved from https://uwspace.uwaterloo.ca/bitstream/handle/10012/6311/Hao_ Yufang.pdf. 22A. Aleman, R. Martin, and W. Ross, “On a theorem of Livsic,” J. Funct. Anal. 264, 999–1048 (2013). 23D. Clark, “One dimensional perturbations of restricted shifts,” J. Anal. Math. 25, 169–191 (1972). 24M. Krein, “On Hermitian operators with deficiency indices one,” in Dokl. Akad. Nauk SSSR, Vol. 43 (1944) pp. 339–342. 25M. Krein, “On one remarkable class of Hermitian operators,” in Dokl. Akad. Nauk SSSR, Vol. 44 (1944) pp. 191–195. 26L. Silva and J. Toloza, “Applications of Krein’s theory of regular symmetric operators to sampling theory,” J. Phys. A 40, 9413 (2007). 27R. Martin, “Representation of symmetric operators with deficiency indices (1, 1) in de Branges space,” Complex Anal. Oper. Theory 5, 545–577 (2011). 28L. deBranges, Hilbert spaces of entire functions (Prentice Hall, 1968). 29M. Gorbachuk and V. Gorbachuk, M.G. Krein’s lectures on entire operators, Vol. 97 (Birkh¨auser,2012). 30S. Garcia, J. Mashreghi, and W. Ross, Introduction to model spaces and their operators, Vol. 148 (Cam- bridge University Press, 2016). 31K. Hoffman, Banach spaces of analytic functions (Courier Corporation, 2007). 32A. Granville and G. Martin, “Prime number races,” The American Mathematical Monthly 113, 1–33 (2006). 33R. J. L. Oliver and K. Soundararajan, “Unexpected biases in the distribution of consecutive primes,” Proceedings of the National Academy of Sciences 113, E4446–E4454 (2016).